Noise recognition of power transformers based on improved MFCC and VQ

Abstract

This paper presents a new method to recognize the noise characteristics of power transformer. First, frame division and windowing are applied to pre-process the noise signals. Then Mel Frequency Cepstrum Coefficient (MFCC) combined with Principal Component Analysis (PCA) is proposed to calculate the feature vectors of noise signals for high accuracy. Finally, the vector quantization (VQ) models are built to recognize the noise characteristics. The noise signals of some 10kV transformer are measured when the core is loosened in different degree. It is shown that the proposed MFCC is capable of describing the noise features of transformer accurately. The results of noise recognition by VQ are agreed well with the preset condition of core. The obtained results are helpful for the optimum design and mechanical condition assessment of power transformer.